Getting Started with Hugging Face Transformers for IPUs with Optimum

Hugging Face · 4 years ago · 9

Hugging Face released Optimum, an open-source library that optimizes Transformer models for faster inference on various hardware platforms, including Graphcore's Intelligence Processing Units. The partnership introduced BERT as the first IPU-optimized model, with the library enabling developers to fine-tune models on IPUs using PyTorch with minimal code changes. Developers can now deploy optimized state-of-the-art models on specialized hardware without requiring extensive optimization expertise, with the example BERT fine-tuning on SQuAD achieving 80.66% exact match accuracy.